Some Buddhists are making a big fuss about the auction of ancient gems that belonged to a man who found them and excavated them. They claim that the gems were mixed with a kind of relic — some of the dust remaining from Buddha's cremation — and this makes them sacred so they can't be auctioned.

I can imagine what the Buddha would say about this obsession:

I invite all of you to learn to be less attached
to physical things, including gems and
derivatives of parts of my former body.
Attachment generates suffering, and that
includes attachment to those things. If my
former body has any importance, it is only in
the advice that it once spoke to you, which I
now repeat.

Please join me in focusing on things that
really matter and you will become more enlightened.

The San Francisco Police Department decided on its own to violate the rules limiting public employee overtime, the Board of Supes learned Tuesday. [...]

At that hearing, some of the supes questioned why SFPD was allowing officers to work as private rent-a-cops, on OT pay, for a lot more than the 500 annual hours that the city currents sets as the limit on extra work.

SFPD also defied a Memorandum of Understanding with the Department of Human Resources that barred cops who had called in sick more than twice in six months to work in the private market, under the so-called 10B program. [...]

Walton said that he's been on the board for seven budget cycles, and every year, the cops have been unable to property estimate their overtime and budget for it. I might be less charitable; one could argue that SFPD doesn't put all its projected OT in the budget because the department brass assume that the supes will just approve supplemental appropriations later. As they have, for seven years.

"We'll never have change if we keep giving them a blank check," Walton said. "We hold everyone accountable except this department. [...]

Then support it they did, 9-2, with only Walton and Fielder in opposition.

Businesses on Market Street, like Salesforce and Twitter, for instance, were frequent customers, as were Union Square stores like Macy's and Victoria's Secret. Officers were also spread out to a handful of retail spots elsewhere in the city, chiefly Walgreens, Target and Safeway.

Walgreens, in fact, was by far the largest user of private security hours from police officers between 2018 and 2023, and many of the blips that appear away from the downtown core are Walgreens stores. (The large bubble in Golden Gate Park? The Outside Lands music festival.) [...]

Officers working private security are considered on-duty and can make arrests. They must wear police uniforms and their badge, and carry department-issued firearms. And they make time-and-a-half: At an officer's starting salary, that's $83.49 an hour. By the time an officer is a sergeant, it's $123.54 an hour, and for a lieutenant, it's $141.06. [...]

Officers might prefer to work private security to shifts in their district stations. The fact that they can easily call out sick and, that very same day, work a private security gig, a practice that was documented extensively in a city audit last year, is a major problem, he said. [...] The program, Walton added, "is a big win for corporations, for private businesses, for overtime abuses, and not for the entire city."

Previously, previously, previously, previously, previously.

(satire) *RFK Jr. Encourages Americans To Do Their Own Research About Dragons.*

British human rights defenders have defeated in court one increase in repression of protests. The increase was serious but the Tories treated it as minor.

I wonder whether this will free any of the UK's political prisoners who were convicted of protesting.

* The saboteur in chief has ordered the closure of 25 scientific centers that monitor US waters for flooding and drought, and manage supply levels to ensure communities around the country don't run out of water.*

From now on, if your region has floods or water shortages, the saboteur will be partly to blame.

While some groups of deportation thugs told mothers being deported that they had to take their US-citizen children with them, another group told another mother that she could not do so.

I conjecture that they have been given vague and confusing instructions, written in a hurry by inexperienced people, so in similar situations various officials draw different conclusions. That tends to happen when those who write the instructions don't think carefully about them.

*{The corrupter's] white-collar criminal pardons cost public $1bn, says ex-DoJ official.*

Assuming he profited from this in some way, it amounts to a transfer from America's account to his account.

*Labour's planning bill threatens protected habitats [in Britain], says environment watchdog.*

There is once again a fleet of unarmed ships carrying aid to Gaza. Someone used drones to attack one of the ships, half the Mediterranean sea away. Attackers suspect the attack was done by Israel.

I agree it was probably the Israeli army , because Israel had means, motive and opportunity. Some other countries had the means, and perhaps an opportunity, but No other country had a motive.

Israel used to seize aid boats and arrest the people on them. The change to attack the boats and maybe kill the people on them reflects how Israel has changed since then.

Was that attack terrorism? I think it was. Terrorism means making war on noncombatants, and those unarmed aid workers are clearly not combatants.

Let’s say that you’re making a deep neural network and want to use toroidal space. For those that don’t know, toroidal space for a given number of dimensions has one value in each dimension between zero and one which ‘wraps around’ so when a value goes above one you subtract one from it and when it goes below zero you add one to it. The distance formula formula in toroidal space is similar to what it is in open-ended space, but instead of the distance in each dimension being a-b it’s that value wrapped around to a value between -1/2 and 1/2, so for example 0.25 stays where it is but 0.75 changes to -0.25 and -0.7 changes to 0.3.

Why would you want to do this? Well, it’s because a variant on toroidal space is probably much better at fitting data than conventional space is for the same number of dimensions. I’ll explain the details of that in a later post1 but it’s similar enough that the techniques for using it an neural network are the same. So I'm going explain in this post how to use toroidal space, even though it’s probably comparable or only slightly better than the conventional approach.

To move from conventional space to an esoteric one you need to define how positions in that space are represented and make analogues of the common operations. Specifically, you need to find analogues for dot product and matrix multiplication and define how back propagation is done across those.

Before we go there it’s necessary to get an intuitive notion of what a vector is and what dot product and matrix multiplication are doing. A vector consists of two things: a distance and a magnitude. A dot product finds the angle between two vectors times their magnitudes. Angle in this case is a type of distance. You might wonder what the intuitive explanation of including the magnitudes is. There isn’t any, you’re better off normalizing for them, known in AI as ‘cosine space’. I’ll just pretend that that’s how it’s always done.

When a vector is multiplied by a matrix, that vector isn’t being treated as a position in space, it’s a list of scalars. Those scalars are each assigned a direction and magnitude of a vector in the matrix. That direction is assigned a weight of the value of the scalar times the magnitude. A weighted average of all the directions is then taken.

The analogue of (normalized) dot product in toroidal space is simply distance. Back propagating over it works how you would expect. There’s a bit of funny business with the possibility of the back propagation causing the values to snap over the 1/2 threshold but the amount of movement is small enough that that’s unusual and AI is so fundamentally handwavy that ignoring things like that doesn’t change the theory much.

The analogue of a matrix in toroidal space is a list of positions and weights. (Unlike in conventional space in toroidal space there’s a type distinction between ‘position’ and ‘position plus weight’ where in conventional space it’s always ‘direction and magnitude’.) To ‘multiply’ a vector by this ‘matrix’ you do a weighted average of all the positions with weights corresponding to the scalar times the given weight. At least, that’s what you would like to do. The problem is that due to the wrap-around nature of space it isn’t clear which image of each position should be used.

To get an intuition for what to do about the multiple images problem, let’s consider the case of only two points. For this case we can find the shortest path between them and simply declare that the weighted average will be along that line segment. If some of the dimensions are close to the 1/2 flip over then either one will at least do something for the other dimensions and there isn’t much signal for that dimension anyway so somewhat noisily using one or the other is fine.

This approach can be generalized to larger numbers of points as follows: First, pick an arbitrary point in space. We’ll think of this as a rough approximation of the eventual solution. Since it’s literally a random point it’s a bad approximation but we’re going to improve that. What we do is find the closest image of each of the points to find a weighted average of to the current approximation and use those positions as the ones when finding the weighted average. That yields a new approximate answer. We then repeat. Most likely in practical circumstances this settles down after only a handful of iterations and if it doesn’t there probably isn’t that much improvement happening with each iteration. There’s an interesting mathematical question as to whether this process must always hit a unique fixed point. I honestly don’t know the answer to that question. If you know the answer please let me know.

The way to back propagate over this operation is to assume that the answer you settled on via the successive approximation process is the ‘right’ one and look at how that one marginally moves with changing the coefficients. As with calculating simple distance the snap-over effects rarely are hit with the small changes involved in individual back propagation adjustments and the propagation doesn’t have to be perfect, it just has to on average produce improvement.

# date Tue 6 May 19:05:57 PDT 2025 # sudo systemctl stop systemd-timesyncd.service # sudo date -s '2232-04-18 16:47:16' date: cannot set date: Invalid argument Wed 18 Apr 16:47:16 PDT 2232 Exit 1 # date Tue 6 May 19:06:14 PDT 2025 # sudo date -s '2232-04-18 16:47:15' Wed 18 Apr 16:47:15 PDT 2232 # sleep 30 ; date Wed 18 Apr 16:47:45 PDT 2232

The cutoff is 0x1ED5D7403.

Previously, previously, previously, previously, previously, previously, previously.

How do I do an Apache rewrite but also have it show up in error_log? E.g. the following has the side effect that it is not logged, making it unavailable to fail2ban:

RewriteCond %{REQUEST_FILENAME} !-f RewriteCond %{REQUEST_FILENAME} !-d RewriteRule ^.*$ - [GONE,LAST]

Previously, previously, previously, previously, previously.

What is your strategy for blocking user agents that are attempting to impersonate browsers?

There are many, many botnets out there sending implausibly old user-agent strings. Now, I'm just an unfrozen caveman, your modern protocols confuse and frighten me. But there's one thing I do know: if I were trying to be stealthy, I would be forging up-to-date UAs.

But, they don't.

I have a long list of UAs that are obviously and 100% illegitimate. Disregard those. It's the UAs that are somewhat plausible that I am asking about.

I'm currently blocking Chrome < 2023, Firefox < 2017, Windows < 2013, macOS < 2013.

If you are here to tell me that:

• People can change their user agent strings;
• Blocking user agents is wrong and bad;
• It is important to cater to the kinks of Vintage Microsoft Windows Retrocomputing Enthusiasts;
• Poor people have old phones, you Imperialist Monster;
• You know what you ought to do, let some third party Clown service MiTM all of your traffic;
• Anything that starts with "Well Actually";

please understand that your opinion has already been noted and will be given all due consideration.

Previously, previously, previously, previously, previously, previously, previously.

The U.N. special rapporteur for the occupied Palestinian territories calls for more European leaders to be charged with complicity in war crimes for supporting Israel's bombardment of Gaza.

*Commerce Secretary Howard Lutnick laid out a disturbing plan to bring back serfdom.*

Community patrol members are specifically trained to look out for signs of undercover law enforcement vehicles -- usually U.S.-manufactured vehicles such as Dodge Durangos, Ford Explorers, and Chevy Impalas with dark-tinted windows that stand out from others in the neighborhood. Trucks and SUVs with no license plates and with lights on the grill and visor are other hallmarks that volunteers look out for.

"There's always an agent inside," Prado said, scanning the streets for suspicious vehicles. "They usually operate in pairs, sometimes you can also see in the backseat, there's like a cell or detention cage." [...]

"Sometimes we'll just ask, 'Are you law enforcement?" he said. "They'll usually just say federal agent. It does give us an indicator that it's some kind of law enforcement. More often than not, we make the assessment to confirm that there is some kind of federal law enforcement, and we'll announce it to the community." [...]

Prado says that the concept of the community-level self-defense was inspired by the Black Panther Party, which formed similar neighborhood patrols to defend their neighborhoods from police terror in the 1960s.

Previously, previously, previously, previously, previously, previously.